The more I learn, the less I know. Working on on-device ML. Building @Private_LLM, @CleanLinksApp and @SlopOrNotAi. ex @google, @facebook
Ireland
- Tweets 2,498
- Following 1,215
- Followers 1,435
- Likes 30,963
Pinned Tweet
If you have a recent-ish Apple device, try out and help me beta test my latest private and fully on-device image generation app.
testflight.apple.com/join/DQ…
1
4
9
5,499
> Ponzi was soliciting investments without figuring out those pesky details of how he would actually make money.
Striking parallels between the OG con artist from over a hundred years ago, and contemporary con artists.
I guess some things never change.
npr.org/2026/10/07/nx-s1-599…
1
89
Jeethu Rao retweeted
We're open-sourcing Rebalancer, the assignment-problem solver Meta has used for resource allocation across our infrastructure for over nine years.
Given a set of objects and a set of bins, Rebalancer assigns one to the other to optimize your objectives under your constraints, including hardware placement, service and task placement, traffic routing, and more. It separates how a problem is specified from how it's stored, solved, and debugged, with both an optimal MIP solver and a parallelized local search solver.
At Meta it solves ~40 million assignment problems a day across 30+ problem formulations. P99 solve time is 12 seconds on a problem with 265k objects and 3.2k bins.
🔗 Read more on our engineering blog: engineering.fb.com/2026/09/2…
42
353
3,710
258,457
1. Burn tons of compute, searching for solutions to well-known problems.
2. Give solutions under NDA to people with high-prestige academic affiliations.
3. Get them to publish it under their own name and affiliations, crediting Claude.
1
1
10
474
Compute bottleneck confirmed.
🚨 BREAKING: Microsoft confirms on publicly accessible web page that OpenAI has been using Looped Transformers in the GPT-6 series, proving The Information's reporting was correct all along‼️
GPT-6.1 Sol uses 2 inference passes, with a passing mention of "instead of three" 👀
4
279
FrogNano-4B-2609 quantized with ParoQuant.
huggingface.co/Jeethu/FrogNa…
5
167
Wow! What a thread?!
now that the dust is settling on @matanSF vs @ScottWu46 , let's do a little post-mortem
the issue is that most people don't understand Cog's business model
2
285
Looks like Codex code reviews have been successfully rugpulled. This was the only use case I had left for the $200 OpenAI sub. I guess it's time to move on.
1
2
190
LoopCD appears to be a training-free inference time trick that could potentially improve generation quality with looped models like @nanbeige Nanbeige4.2-3B.
Very reminiscent of Contrastive Decoding (arXiv:2309.09117) and CFG from diffusion models. Although unlike the former two, LoopCD-Hidden is a free lunch.
Can't wait to try it out!
Introducing LoopCD our latest research at 🍎 Apple MLR.
Heard the rumors that frontier models like GPT-6 Astra and Gemini 4 use Looped Transformers? We present a new inference-time method that makes pre-trained looped models better for (almost) free!
AIME 2024 (Ouro-2.6B): 61.9 → 73.3
HumanEval (Huginn): 22.6 → 31.7
No training. No extra model.
Read our paper for more details: arxiv.org/abs/2610.02185
1
2
5
508
4-bit learned quants are excellent, often near lossless. It’s low-effort affine (aka RTN) quants that are the problem.
MLX has DWQ (learned scales and biases), which is quite good. If you want to see how good 4-bit quants can get, you might want to look at QAT/QAD checkpoints.
Please stop using MLX 4bit affine quants, focused on something else. They are fast but not good enough, especially in larger contexts.
I learned the lesson the hard way using them and even thanks to @ggerganov pointing me in the right direction a long time ago.
3
1
7
909
Looks like this account is run by perfidious marketers. Both the speed and accuracy numbers are fabricated. 😂
Run decision models with our llama.cpp engine ⚙️
We're bringing support for over 10 System One models into Atomic Chat, now your device makes choices 10.43x faster than Jev's API utilising only the CPU
Run decision models locally -> atomic.chat
2
2
5
263
I'd recommend reading this cloudflare blogpost, if you're looking for sound benchmarks, rather than straight up lies.
blog.cloudflare.com/clef-dec…
1
34
Let the gatekeeping begin.
Hey David, Gayani here from Figma. You're right that our remote MCP server only accepts clients on our supported list, and Pi isn't on it yet. You can see the current list in our MCP catalog at figma.com/mcp-catalog. If you'd like Pi considered for a future addition, please fill out this form: forms.gle/qSvUawwznWyoj8Go7.
2
109
People who’re into counting/maxing tokens today are the same people who used to count/max vanity metrics like lines of code and git commits in the previous era.
2
2
7
305
Can't believe they managed to land a plane with a missing rudder. That's some ace piloting!
Severe damage to the aircraft involved in the FlyDubai incident, a Boeing 737 MAX 8, following the ‘forced’ descent (from 30,000ft to around 15,000ft in just approx 30 secs) during the time of an incident in the flight deck (involving pilots, and reportedly some passengers)
1
197
Jeethu Rao retweeted
If you have a recent-ish Apple device, try out and help me beta test my latest private and fully on-device image generation app.
testflight.apple.com/join/DQ…
1
4
9
5,499
If you have a recent-ish Apple device, try out and help me beta test my latest private and fully on-device image generation app.
testflight.apple.com/join/DQ…
1
4
9
5,499
New family of open Tabular Foundation Models from NVIDIA. Three sizes for classification and regression: Small, Medium and Large.
I’m very excited to share @nvidia Kumo Tabular, a new family of foundation models for tabular data.
Kumo Tabular establishes the new Pareto frontier across the entire accuracy–inference-time tradeoff.
Just as importantly, we are releasing it openly: open weights, open-source software, and a permissive license for commercial use.
HuggingFace: huggingface.co/nvidia/Kumo-T…
GitHub: github.com/NVIDIA/structured…
1
1
9
1,710